CSE DSI Machine Learning Seminar - Haruka Kiyohara (Computer Science, Cornell)

End-to-End Training of Two-Stage Decision Systems: Towards Personalized Decisions at Scale

Modern decision-making systems, including e-commerce, search, chatbots, and social networking feeds, need to handle massive volumes of items at web latency. For this reason, typical recommender and retrieval-augment generation (RAG) systems employ two-stage architectures; an early-stage model that focuses on inference speed performing coarse-grained decisions, and a late-stage model focusing on model expressiveness performing fine-grained decisions. An effective early-stage model is crucial in this two-stage pipeline, as it serves as the performance bottleneck. However, efficient training methods for the early-stage model have been underexplored.

In this talk, we discuss a data-efficient approach for training the early-stage decision model end-to-end using the user-provided implicit feedback such as clicks or purchases. I will also present several open challenges for early-stage decision learning, and explore how to achieve a better tradeoff between inference latency and model flexibility for large-scale applications.

Haruka Kiyohara is a fourth-year Computer Science Ph.D. candidate at Cornell University. Her research interest lies in evaluating and optimizing decision-making systems using causal inference and machine learning, particularly learning from logged data and optimizing for long-term social goods in large-scale recommender systems. Her work has been published at machine learning and data mining conferences, including ICML, NeurIPS, ICLR, KDD, WSDM, and RecSys. Prior to Cornell, she received a B.E. in Industrial Engineering and Economics from Tokyo Institute of Technology with the Excellent Student Award. Her Ph.D. study is supported by the Funai Overseas Scholarship, Quad Fellowship, and a gift to the LinkedIn-Cornell Bowers Strategic Partnership. In her free time, she enjoys creating songs collaborating with generative AI tools.